AI Transparency

Modulon uses artificial intelligence to take routine work off a clinic's team: answering calls and texts, booking, reading the inbox, drafting notes and preparing billing work. This page explains where the AI is used, what it may and may not do, the models behind it, and how clinic staff and patients stay in charge of it.

Our approach rests on one principle: the clinic's rules decide, not the AI. Rules are enforced in software, so the AI cannot act outside them, and anything clinical waits for a clinician.

• Diagnose, triage, or give medical advice. Clinical questions go to the clinic's staff.
• Send results, refills or any other clinical message without a clinician's approval.
• Act outside the rules the clinic has set, such as who may be booked, when, and with whom.
• Deny anyone care or coverage. Where a request cannot be met, it hands the decision to a person.
• Pretend to be a person. It says it is an automated assistant.
• Handle an emergency. It tells the caller to contact emergency services and alerts the clinic.

• Calls answered by Modulon start by saying the caller is speaking with an automated assistant, and that the call is recorded.
• The first text in a conversation says it comes from an automated assistant for the clinic.
• Patients can ask for a person at any time, and the assistant transfers them or arranges a call back.
• The AI note taker records a visit only after the patient agrees, and the provider can switch it off at any point.
• Messages drafted by the AI that carry clinical information are reviewed by a licensed clinician before they are sent.

Clinics can add their own wording to these disclosures, but cannot switch them off.

Rules in software

Each clinic writes down its rules with us during onboarding: which visit types may be booked, with which providers, how many new patients a day, quiet hours for texts, and when to hand off. The platform checks every action against them before it happens.

Watch first, then one task at a time

New tasks run in a watch-only mode first: the AI drafts what it would do, and sends nothing. The clinic turns each task on only after it has seen the drafts, and can turn it off again at any time.

A record of every action

Every action the AI takes is recorded with the rule that allowed it, so staff can see what it did and why, correct it, and tell us when it got something wrong.

Modulon is built on large language and speech models from established providers, run in the clinic's region:

• In the United States, the European Union and the United Kingdom, models from OpenAI.
• In the Gulf, the same models through Microsoft's Azure OpenAI Service, run in the region.

Every provider processes patient data under a Business Associate Agreement or Data Processing Agreement, keeps none of it, and does not use it to train its models. We do not use patient data to train models either. The providers in each region are listed on our subprocessors page.

AI can be wrong, and we design the product around the ways it can fail:

• It can mishear names, numbers and medication names, especially on poor lines, and in strong accents or regional dialects of English and Arabic. It reads details back to confirm them, and flags low confidence.
• It can misread handwriting and poor scans. Low-confidence fields are sent to staff rather than filed.
• A drafted note can include something that was not said or miss something that was. That is why the provider reviews and signs every note.
• It can misunderstand an unusual request. When it is unsure, it hands off rather than guessing.

• Before each release, we test the AI against a library of real-world cases, with patient details removed, for accuracy, rule-following and handoffs.
• We test it against misuse: attempts to talk it out of a rule, to extract another patient's information, or to make it give medical advice.
• We test it in each language and accent we support, and check that its handling does not differ unfairly between groups of patients.
• In production, we monitor handoff rates, corrections by staff and patient complaints, and investigate any drift. Serious errors are handled through our incident response plan.

We design the AI to meet the rules on AI in each region we serve, and help clinics meet their own:

• European Union. The EU AI Act's transparency duties for systems that talk with people. Modulon is not a medical device and does not triage emergencies, so it is not a high-risk system under the Act.
• United States. State laws on AI in patient communications and healthcare, such as California's rules for AI-generated clinical messages, Utah's and Texas's disclosure rules, and Colorado's rules on consequential decisions, which the AI does not make.
• United Kingdom. The ICO's guidance on AI and data protection.
• Gulf. The AI policies of the Department of Health Abu Dhabi and the Dubai Health Authority, the DIFC's rules on autonomous and AI systems, and SDAIA's AI Ethics Principles in Saudi Arabia.

We keep the documentation clinics need for their own assessments, and share it on request.

If you have a question about how Modulon uses AI, or believe it got something wrong, write to hello@modulonhealth.com. Patients can also raise concerns with their clinic, which can review the record of every action the AI took.